Print
Discuss
Send to friend
RSS feedsNo matter how well planned and managed, there will come a time when you need to consider expanding or upgrading your network. Typically that's because it no longer provides the kind of performance, capacity and reliability needed to support your business, its users or the applications and data they need.
Unfortunately upgrading is a far from easy task, with lots of hardware and software being launched every week that could potentially be of value. The real art, therefore, comes in knowing what to upgrade for maximum benefit, at the lowest cost in terms of both money and effort.
You don't, for example, want to spend thousands of pounds on new servers when simply putting extra memory or adaptors into existing systems would have the same effect.
Similarly, a change in software might appear cheaper than revamping relatively new hardware, but you have to factor in the disruption caused and any additional training required.
In this feature we'll be discussing some of the most common upgrades and their relative benefits, such as installing new local area network (Lan) switches, adding server processors and memory, fine tuning network storage and upgrading to the latest operating system software.
Where possible we'll also use real results from recent VNU Labs tests to show you what effect such changes might have. We can't tell you exactly what to upgrade, or in what order (that all depends on where you're starting from and what you want to achieve) but you'll get a good idea of what you might want to try first and what you can expect to get in return.
Expanding the infrastructure
One of the first elements to look at is the network infrastructure, in terms of the cables, hubs, switches, routers and other devices. Ethernet hubs are old hat and changing to switches will make better use of available bandwidth and provide immediate performance gains for little money or effort.
By replacing older 10Mbps Ethernet hubs and switches with hardware that supports Fast (100Mbps) or Gigabit (1,000Mbps) Ethernet, users will see huge improvements in the speed of the Lan.
Bear in mind, however, that it's no good having switches capable of supporting 100Mbps or Gigabit speeds to the desktop if those desktops aren't equipped with similarly specified Lan adaptors. On top of that, old network cabling may have to be upgraded to handle the faster speeds, especially Gigabit Ethernet, which needs good-quality Category 5 cabling or above.
If your network infrastructure has grown and been enhanced over the years, it's worth taking the time to review how the plethora of devices fit together. An assortment of distributed switches may do the job, but they aren't necessarily a good idea as they use up network bandwidth communicating with each other.
Stackable products, with independent communication links, are more expensive but could be worth buying if they have less impact on overall network performance. Likewise, chassis-based switches, with dedicated backplanes, offer simple expansion and management facilities and are a better investment on large networks.
Consider using managed hardware to better control network performance and availability. On a large network it's worth investigating more intelligent switches that can look at packet information at Layer 3 and above to prioritise and route traffic, depending on the users and applications involved.
Compared to ordinary Layer 2 switches these are a lot more expensive but, again, can provide large performance gains with minimal effort or disruption to the network and your business.
Server connectivity
Try and avoid bottlenecks where shared resources, such as servers, connect to the network. Most servers come with just one network interface, but there's no reason not to put in extra Lan adaptors or upgrade to faster interfaces.
Simply fitting a second Lan adaptor can dramatically improve throughput on hard-pressed servers, especially those used mainly for file sharing. In our tests we saw improvements of over 30 per cent without other changes being made. Adding yet more adaptors may provide further gains, although extra adaptors may have an adverse impact unless other components are also upgraded.
Although extra Lan adaptors provide extra network bandwidth, enabling more data to be sent to and from the server, each adaptor is also a drain on CPU resources. The entry-level server in our tests couldn't handle more than two adaptors effectively, but servers with faster processors can take more.
It's also possible to get Lan adaptors designed expressly for server use, with several Ethernet ports on one card and onboard processors to offload much of the work needed to handle TCP/IP communications.
These are becoming increasingly common, especially with the introduction of IP-based storage area networks (Sans), which rely on fast Ethernet connectivity to work.
Another approach is to upgrade to fewer, faster adaptors, although the impact of this change may not be as great as the claimed data rates might lead you to believe. For example, we replaced the 10/100 Ethernet adaptors with a single Gigabit Ethernet card.
Compared to adding a third 10/100 card there was a significant improvement but it wasn't the five-fold rise you might otherwise expect when going from 200 to 1,000Mbps. Other factors, such as processing overheads, available memory and the capabilities of the storage subsystem also have an effect and need to be taken into consideration.
Two Gigabit adaptors would be even better and, when upgrading to faster Lan adaptors it's worth equipping servers with two or more interfaces for extra redundancy.
However, the mapping of DNS and Netbios names to multiple IP addresses is complex. You could easily end up with one adaptor doing all the work and the others sitting idle. Most adaptors, therefore, come with 'teaming' or 'load balancing' drivers which can be used to map multiple interfaces to a single IP address.
This makes for simpler management of naming services and helps spread loads more evenly, in some cases taking account of demand conditions, user priorities and other variables. The load balancing software can also handle automatic failover of connections in the event of a fault, providing useful extra resilience on top of the performance gains.
More processing power
The next thing to look at is processing power although, as with Lan adaptors, this isn't straightforward. Common sense says that the faster the processor the higher the level of performance you're likely to get, and in desktop PCs that's generally the case.
In a server, however, the processor may not be doing as much as you might expect, and switching to something faster may not always boost performance. This is especially the case when the prime use is file and printer sharing.
In such instances upgrading from a Pentium III to a Pentium 4, Xeon or even a 64bit Itanium 2 server is, by itself, likely to have only a small effect on throughput. And, given that you normally have to swap out the whole server to make such a change, you could be wasting a lot of money and disrupting the network for little gain.
Upgrading a symmetrical multiprocessing (SMP) server, where two or more processors share the workload, is a lot easier, but the same considerations apply.
We tested a Xeon server fitted with one and two processors running a typical light mix of file sharing tasks. At low client loads the second processor did provide a small boost but, as the number of clients increased, the overheads associated with SMP depressed performance compared to a single CPU.
In practice that means word processing documents, emails and spreadsheets are unlikely to open or save any quicker on a server with faster or more processors. Printing will also take just as long. You might notice an effect with large numbers of users or if serving up big multimedia files, but upgrades to other components could provide the same, if not better results.
The same doesn't apply when the server is used as a platform for database, web and other applications. In such cases the processor can be a major determinant of performance and increasing the speed or adding more processors really can have an effect.
We tested a Xeon-based Windows 2000 server fitted with one, two, three and four processors. The results showed impressive gains as the extra processors are added, but there is a significant cost.
The Xeon DP/MP and Itanium-2 processors used in scalable mid-range servers, for example, aren't cheap and, depending on speed and cache size, it can cost several thousand pounds to upgrade to a full complement.
Expansion beyond four-way is particularly expensive and above four CPUs the benefits diminish anyway, making such configurations of little value unless you're running really demanding database applications that need to be hosted on a single server.
On the plus side, adding extra processors causes very little disruption or downtime other than that needed to fit the new chips. It's also a good way of consolidating server hardware to simplify maintenance and management.
There's better news lower down the scale, where more modest Pentium 4 and Xeon processors aren't nearly so costly. Most entry-level servers come with two-processor sockets as standard, making an extra CPU an easy as well as a cheap upgrade should you need the additional power.
That is something that's easily checked in Windows using the Performance Monitor tool, which shows how much time the processors spend working. If you're regularly seeing high CPU usage (80 per cent and above) another processor might well help.
The latest Intel chips also offer support for a technology called hyper-threading. As in an SMP server, this allows application code to be broken down into separate threads that are then run concurrently. However, whereas with SMP this is done with two or more processors, with hyper-threading just one chip is needed.
On servers fitted with recent P4s and Xeons the operating system sees two logical hyper-threaded processors for each physical chip installed. A single processor, therefore, appears to be a dual-processing system, two chips look like a four-way server and so on.
The operating system code, which could be Windows, Linux or any MPS-aware software, and any applications run as normal, with only minor modifications needed for developers to take full advantage of the technology.
As with faster or multiple processors, hyper-threading is of little value when it comes to file and print sharing. But it does have an effect when it comes to application hosting. We tested a server with and without hyper-threading. Turning on the technology in the Bios provided an immediate performance boost, although at high client loads the effect was again minimised as other factors came into play.
Not all servers can be upgraded to processors with hyper-threading, especially older systems, and software needs to be written with multithreading in mind to take full advantage. For maximum benefit you might also need an operating system upgrade.
Memory effects
As well as speed, cache memory size is an important factor to consider when upgrading server processors. Older chips, typically, will have very little of the onboard cache used to speed up memory access.
Anything from 64KB to 512KB of Level 2 cache is usual on Pentium II and III chips, but that has changed with several megabytes now available on some of the more recent Intel products. Xeon MP processors, for example, ship with up to 512KB of Level 2 cache, together with up to 2MB of Level 3 cache onboard.
Higher up the scale, the 64bit Itanium 2 can be specified with a whopping 6MB of Level 3 cache and in all cases the extra cache memory can make a big difference when it comes to application performance, as can extra memory in general.
Memory is no longer the frighteningly expensive commodity it once was and there's little point skimping on what you install. Even on an entry-level file and print server there's not much to be saved by specifying less than 512MB.
Not that much of this is needed for processing duties; the extra Ram is used by the network operating system (Windows, Netware, Linux etc) for data caching when sharing out files. The software does this automatically, moving large blocks of disk data into memory even though it may only be transferring small amounts to clients at any one time.
With more of the data in memory, chances are the next request can be satisfied without having to go back to the disk itself, thus speeding up access. Similarly, file updates can be cached in memory and flushed to the disk less often.
To see the effect of extra memory on file server throughput on a modest Windows server we doubled the amount of memory on an old dual Pentium III server, originally supplied with 256MB of Ram, improving file sharing performance by over 25 per cent at high loads.
Upgrading to 1GB gave further gains, but the improvement wasn't anywhere near as large. This was because the server we tested was originally fitted with just one 10/100Mbps Lan interface, which limited the number of client requests it was able to handle.
When we fitted a second adaptor the extra memory beyond 512MB had a much bigger effect, illustrating the importance of considering changes to other components when embarking on any upgrade.
See also:
How faster networking technology can streamline processes and increase productivity 31 Jul 2004
Has Windows Server 2003 lived up to Microsoft's claims of being the 'most reliable, highest-performing server operating system' the company has ever built? 12 May 2004All Home Networks
del.icio.us
Digg this
reddit!
